CN110596645A - Two-dimensional inversion-free sparse Bayesian learning rapid sparse reconstruction method - Google Patents

Two-dimensional inversion-free sparse Bayesian learning rapid sparse reconstruction method Download PDF

Info

Publication number
CN110596645A
CN110596645A CN201910854202.5A CN201910854202A CN110596645A CN 110596645 A CN110596645 A CN 110596645A CN 201910854202 A CN201910854202 A CN 201910854202A CN 110596645 A CN110596645 A CN 110596645A
Authority
CN
China
Prior art keywords
sparse
matrix
vectorized
signal
noise
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Granted
Application number
CN201910854202.5A
Other languages
Chinese (zh)
Other versions
CN110596645B (en
Inventor
张双辉
刘永祥
黎湘
霍凯
姜卫东
高勋章
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
National University of Defense Technology
Original Assignee
National University of Defense Technology
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by National University of Defense Technology filed Critical National University of Defense Technology
Priority to CN201910854202.5A priority Critical patent/CN110596645B/en
Publication of CN110596645A publication Critical patent/CN110596645A/en
Application granted granted Critical
Publication of CN110596645B publication Critical patent/CN110596645B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Classifications

    • G—PHYSICS
    • G01—MEASURING; TESTING
    • G01S—RADIO DIRECTION-FINDING; RADIO NAVIGATION; DETERMINING DISTANCE OR VELOCITY BY USE OF RADIO WAVES; LOCATING OR PRESENCE-DETECTING BY USE OF THE REFLECTION OR RERADIATION OF RADIO WAVES; ANALOGOUS ARRANGEMENTS USING OTHER WAVES
    • G01S7/00—Details of systems according to groups G01S13/00, G01S15/00, G01S17/00
    • G01S7/02—Details of systems according to groups G01S13/00, G01S15/00, G01S17/00 of systems according to group G01S13/00

Abstract

The invention belongs to the field of signal processing, and particularly relates to a two-dimensional inversion-free sparse Bayesian learning rapid sparse reconstruction method, which comprises the following steps of: s1: carrying out sparse representation modeling on a two-dimensional sparse reconstruction problem; s2: carrying out statistical modeling on a vectorization sparse signal x and vectorization noise n; s3: solving the posterior probability of the vectorized sparse signal x, the vectorized inverse variance gamma and the noise inverse variance alpha; s4: the matrix form Z of the auxiliary variables is updated. Compared with the IFSBL method, the method disclosed by the invention directly processes the two-dimensional signals, avoids the problem that the two-dimensional signals are vectorized to generate a large matrix, obviously improves the operation efficiency and obviously reduces the requirement on a calculation memory; on the other hand, sparse reconstruction is realized under a statistical signal processing framework, compared with a non-statistical sparse reconstruction method, the method has the advantages of being easy to obtain a global optimal solution, strong in noise robustness, low in dependence degree of algorithm performance on parameter initialization and the like, and the engineering practicability is high.

Description

Two-dimensional inversion-free sparse Bayesian learning rapid sparse reconstruction method
Technical Field
The invention belongs to the field of signal processing, and particularly relates to a two-dimensional inversion-Free Sparse Bayesian Learning (2D-IFSBL) fast Sparse reconstruction method.
Background
Sparse reconstruction is the core of the compressed sensing technology, and a sparse signal can be accurately reconstructed from incomplete observation data. With continuous development, sparse reconstruction technology has been widely applied in the fields of medical image processing, computer vision, radar imaging, and the like. Classical Sparse reconstruction algorithms include l1 regularization, Basis Pursuit (BP), Orthogonal Matching Pursuit (OMP), Sparse Bayesian Learning (SBL), and the like. The SBL method solves the problem of sparse reconstruction under the framework of a statistical theory, performs statistical prior modeling and posterior solution on an observation signal, a sparse signal and a noise signal, and1compared with methods such as regularization, BP, OMP and the like, the method has the advantages of being easy to obtain global optimal solution, strong in noise robustness, self-learning of parameters and the like. However, in the solving process of the SBL method, complex matrix inversion operation is involved, and the calculation complexity is high. Literature (Duan H, Yang L, Fang J, and Li H. fast Inverse-Free spark Bayesian Learning via modified evolution [ J.]IEEE Signal Processing Letters,2017,24(6): 774-.
In practical application, a two-dimensional signal sparse reconstruction problem exists, for example, in radar imaging, two-dimensional Fourier transform needs to be carried out on radar echoes, and a two-dimensional sparse reconstruction problem is to reconstruct a radar image from an incomplete radar echo. When the problems are processed, the traditional processing flow is to stack the two-dimensional signals according to columns for vectorization and then process the signals by using the traditional sparse reconstruction method, but a matrix with a larger size is introduced at the moment, so that the operation efficiency is low, the requirement on the memory is high, and the method has the advantages of low operation efficiency and low memory requirement1The traditional methods such as regularization, BP, OMP, SBL and the like are not satisfiedAnd (5) actual engineering requirements. Although the IFSBL method improves the operation efficiency compared with the SBL method, the two-dimensional signal also needs to be vectorized when the two-dimensional signal is processed, so that large matrix operation cannot be avoided, and the operation efficiency still cannot meet the actual engineering requirements.
Disclosure of Invention
The invention aims to solve the technical problem of how to reconstruct two-dimensional sparse signals from incomplete observation data quickly and efficiently, and reduce the requirement on operation memory so as to improve the engineering practicability.
The invention provides a two-dimensional inversion-free sparse Bayesian learning method aiming at the problem of low operation efficiency of a traditional sparse reconstruction method for processing two-dimensional signals. The method comprises the steps of firstly carrying out sparse representation modeling on a two-dimensional sparse reconstruction problem, and respectively carrying out statistic prior modeling on a sparse signal to be reconstructed and noise; and then calculating the posterior probability of the sparse signal through a Bayes formula based on the prior of the sparse signal to be reconstructed and the likelihood function of the observation signal, thereby realizing the reconstruction of the sparse signal. The two-dimensional signals are directly processed in the calculation process without vectorization, and large matrix processing is avoided, so that the operation efficiency of two-dimensional sparse reconstruction is effectively improved, and the memory requirement is reduced.
The technical scheme adopted by the invention for solving the technical problems is as follows: a two-dimensional inversion-free sparse Bayesian learning rapid sparse reconstruction method comprises the following steps:
s1 sparse representation modeling is carried out on the two-dimensional sparse reconstruction problem:
reconstructing a two-dimensional sparse signal from incomplete observation data, firstly, establishing a relational expression of the two signals, as shown in the following formula:
Y＝AXBT+N (1)
whereinRespectively representing an observed signal, a left dictionary matrix, a sparse signal, a right dictionary matrix and noise,representing a set of real numbers, P, Q, M, N representing the dimensions of the matrix, respectively. In the two-dimensional sparse reconstruction problem, the observed signal dimension is smaller than the sparse signal, i.e. P < M and Q < N. Because the observation signal and the sparse signal are two-dimensional signals, the sparse signal X cannot be reconstructed from the observation signal Y directly by adopting a traditional sparse reconstruction method, and firstly, the observation signal and the sparse signal need to be vectorized:
y＝Φx+n (2)
wherein y ═ vec (y), x ═ vec (x), n ═ vec (n) respectively represent vectorization of observed signals, sparse signals, and noise, vec (·) represents vectorization of the matrix by column stacking;the representation matrix vectorizes the corresponding dictionary matrix:whereinRepresenting the kronecker product of the matrix. The dimension of the dictionary matrix phi is large, so that the problem shown in the formula (2) solved by the traditional sparse reconstruction method is low in operation efficiency, and the requirement on the memory is high, for example, when the sizes of the observation signal Y and the sparse signal X are respectively 64 × 64 and 128 × 128, the size of the dictionary matrix phi is 8192 × 8192, and the efficiency of directly processing the dictionary matrix phi cannot meet the engineering requirement.
S2 statistically models the vectorized sparse signal x and the vectorized noise n:
in the statistical modeling process, the Gaussian distribution of a two-layer structure is adopted to carry out statistical modeling on the quantitative sparse signal x. In the first layer, it is assumed that the elements of the vectorized sparse signal x obey a gaussian distribution with a mean value of zero and independent of each other:
where γ represents the inverse variance of the vectorization, xn、γnSeparately representing vectorized sparse signals x and vectorsThe nth element of the inverse variance γ is normalized.
In the second layer, the vectorized inverse variance γ is assumed to obey the gamma distribution:
wherein a and b represent the shape parameter and the size parameter of the gamma distribution, respectively.
The statistical modeling of the quantization noise n is also performed using a two-layer structure gaussian distribution. In the first layer, assuming that the vectorized noise n obeys a zero-mean gaussian distribution, the likelihood function of the vectorized observed signal y also obeys a gaussian distribution:
whereinDenotes a gaussian distribution, I denotes an identity matrix, and α denotes the inverse of the noise variance.
In the second layer, it is assumed that the inverse α of the noise variance obeys the gamma distribution:
wherein c and d respectively represent the shape parameter and the scale parameter of the gamma distribution.
S3 solves the posterior probabilities of the vectorized sparse signal x, the vectorized inverse variance γ, and the noise inverse variance α:
the process of realizing sparse reconstruction by the SBL method is essentially a process of deducing the posterior probability of the sparse signal by a Bayes formula based on a likelihood function of an observation signal and a prior model of the sparse signal, while the traditional SBL method involves a complex matrix inversion process when solving the vectorized sparse signal x posterior probability, and has low operation efficiency. Matrix inversion is effectively avoided by an inversion-Free Sparse Bayesian learning method (IFSBL) (Duan H, Yang L, Fang J, and Li H. fast Inverse-Free Sparse learning via delayed inversion of evaluation Low Bound approximation [ J ]. IEEE Signal Processing Letters,2017,24(6): 774-. Specifically, the IFSBL method is based on a likelihood function of the vectorized observation signal y shown in equation (5) and a prior probability of the vectorized sparse signal x shown in equation (3), and the posterior probability of the vectorized sparse signal x is obtained as follows:
where the desired μ and covariance matrix Σ is given by:
whereinAs an auxiliary variable, the number of variables,<·>expressing the calculation of an expected operator, wherein T is a constant and takes a value slightly larger than a Lipschitz constant: t2 lambdamax(ΦTΦ)+10-10Wherein λ ismax(ΦTPhi) represents phiTMaximum eigenvalue of Φ. diag (·) denotes the generation of a diagonal matrix using vectors. As can be seen from equation (9), the covariance matrix Σ is a diagonal matrix, and thus the inversion thereof can be obtained by inverting each element, and the operation efficiency is significantly improved. However, the calculation of the large matrix Φ in equation (8) is inefficient and requires a high memory. Aiming at the problem, the invention corrects the expectation solving process shown in the formula (8) so that the expectation solving process can directly process two-dimensional data to avoid large matrix calculation caused by matrix vectorization. Specifically, formula (9) can be substituted for formula (8):
whereinRepresenting the division of the elements of the matrix, 1MN×1A vector having a size of MN × 1 and all elements of 1 is represented. Further will beSubstituting formula (10) and utilizing the relevant properties of the kronecker product, one can obtain:
the above equation can be transformed into a matrix form:
whereinIn the form of matrices of the desired mu, the auxiliary variable z and the inverse of the vectorized variance gamma, 1M×NIs represented by 1MN×1In the form of a matrix. As shown in equation (12), the calculation process of U only involves matrix operation with the maximum size of M × N, so the calculation efficiency is significantly higher than that of equation (8), and the requirement on the memory is reduced.
The posterior probability of the matrix form Γ of the vectorized inverse variance γ is further calculated, which is obtained by means of the bayesian formula, which also obeys the gamma distribution:
wherein the shape parameterAnd scale parameterAs shown in the following formula:
wherein Xm,nThe m-th row and n-th column elements representing the sparse signal X,is expected toThe calculation expression of (a) is:wherein U ism,n、Γm,nThe m-th row and n-th column elements of U and Γ are shown, respectively.
And finally, calculating the posterior probability of the inverse noise variance alpha, which is known by a Bayesian formula and also obeys gamma distribution:
wherein the shape parameter of the a posteriori probabilityAnd a scale parameterRespectively as follows:
wherein sum (·),. fwdarw respectively denote the addition of elements in the matrix and the multiplication of each element of the two matrices.
The desired < α > and < Γ > values referred to in equations (12), (18) may be obtained from the posterior probabilities shown in equations (16) and (13), respectively, as shown in the following equations:
<Γm,nmatrix form representing vectorized inverse variance γ expectation<Expectation of each element in gamma;
s4 updates the matrix form Z of the auxiliary variable:
after the posterior probabilities of the vectorized sparse signal x, the vectorized inverse variance γ and the noise inverse variance α are calculated, the matrix form Z of the auxiliary variable is further updated by the following formula:
Z＝U (21)
in summary, the process flow of the present invention can be summarized as follows: and (3) circularly iterating the formula (12), the formula (19), the formula (20) and the formula (21) until the relative error between the U obtained by two adjacent iterations reaches a rated threshold (for example | | U:)(k+1)-U(k)||2/||U(k)||2＜10-3Wherein U is(k)、U(k+1)Respectively representing the U obtained by the k < th > and k +1 < th > iterations, | | · | | luminous flux2Representing 2 norm of the matrix), and the U obtained in the last step is the reconstructed sparse signal. Furthermore, the parameters need to be initialized before the iterative process, in particular: u is initialized toWhereinRepresenting a pseudo-inverse of the matrix; the inverse of the noise variance, α, is initialized to α ═ 0.1var (y)-1Wherein var (·) represents a variance operator; the parameters a, b, c, and d in the gamma distributions shown in equations (4) and (6) are initialized to a, b, c, d, and 10-6(ii) a The vectorization inverse variance matrix form gamma is initialized to be 1M×N。
The invention has the following beneficial effects: compared with the IFSBL method, the method can directly process the two-dimensional signals, avoids the problem that the two-dimensional signals are vectorized to generate a large matrix, obviously improves the operation efficiency and obviously reduces the requirement on a calculation memory; on the other hand, the 2D-IFSBL method provided by the invention realizes sparse reconstruction under a statistical signal processing framework, and has the advantages of being easier to obtain a global optimal solution, stronger in noise robustness, lower in degree of dependence of algorithm performance on parameter initialization and the like compared with a non-statistical sparse reconstruction method, and the engineering practicability is strong.
Drawings
FIG. 1 is a flow chart of an embodiment of the present invention;
FIG. 2 shows sparse reconstruction results under different observation signal sizes;
FIG. 3 shows sparse reconstruction results under different signal-to-noise ratio (SNR) conditions;
FIG. 4 is a comparison of relative estimation errors of sparse signals reconstructed by different methods under different observation signal sizes;
FIG. 5 is a comparison of relative estimation errors of sparse signals reconstructed by different methods under different SNR conditions;
FIG. 6 shows a time comparison calculated by different methods for different observed signal sizes.
Detailed Description
The invention is further illustrated with reference to the accompanying drawings:
FIG. 1 is a general process flow of the present invention.
The invention relates to a two-dimensional inversion-free sparse Bayesian learning rapid sparse reconstruction method, which comprises the following four steps of:
s1: carrying out sparse representation modeling on a two-dimensional sparse reconstruction problem;
s2: carrying out statistical modeling on a vectorization sparse signal x and vectorization noise n;
s3: solving the posterior probability of the vectorized sparse signal x, the vectorized inverse variance gamma and the noise inverse variance alpha;
s4: the matrix form Z of the auxiliary variables is updated.
Experiments were conducted using simulation data to compare the performance of the method of the present invention with IFSBL method (Duan H, Yang L, Fang J, and Li H. fast Inverse-Free spark Bayesian Learning via delayed evolution lower Bound maximum knowledge [ J ]. IEEE Signal Processing Letters,2017,24(6): 774-) 778) and two-dimensional smooth 0-norm method (2D-SL0) (Aboozar G, Massoud B Z, Christian J. spark analysis of two-dimensional signals [ C ]. IEEE International Conference on Acoustics, Speech and Signal Processing 2009, Taipei 2009: 3157-. In the experimental process, firstly, a two-dimensional sparse signal X with the size of M multiplied by N is generated in a simulation mode, and the signal comprises K non-zero points; the sizes of the left dictionary matrix A and the right dictionary matrix B are respectively P multiplied by M, Q multiplied by N, and all elements of the left dictionary matrix A and the right dictionary matrix B are subjected to standard normal distribution; the noise N is white Gaussian noise; finally, the observation signal Y is generated by equation (1).
In the first experiment, it is assumed that the two-dimensional sparse signal X includes K — 9 non-zero points, each of which takes a value of 1 and is uniformly distributed in the center of the image, so as to facilitate comparison. The size of the two-dimensional sparse signal X is set to 80 × 80, i.e., M ═ N ═ 80. The signal-to-noise ratio (SNR) of the observation signal Y is set to 20 dB. Changing the size of the observed signal Y: let P-Q25, 40, 55 to compare the performance of the three methods 2D-SL0, IFSBL and the 2D-IFSBL proposed by the present invention under different observed signal size conditions. Fig. 2 shows the sparse signals reconstructed by the three methods, where the first, second, and third rows correspond to the sparse reconstruction results under the conditions of P25, 40, and 55, respectively. As can be seen from the figure, the sparse signals obtained by the 2D-IFSBL and IFSBL methods provided by the present invention are completely consistent, and when P is 25, the 2D-SL0 reconstruction result is better, and when P is 55, the results obtained by the 2D-IFSBL and IFSBL methods provided by the present invention are less affected by noise, and the reconstructed sparse signal is better than the 2D-SL0 method.
The second experiment compares the performance of the three methods under different SNR conditions. The experimental parameters were set as follows: P-Q-55, M-N-80, K-9, SNR 15dB, 7.5dB, and 0dB respectively, and the sparse signals obtained by the three methods under different SNR conditions are shown in fig. 3. As can be seen from the figure, the sparse signals reconstructed by the 2D-IFSBL and IFSBL methods provided by the invention are obviously affected by noise to a lower extent than the 2D-SL0 method, especially under the condition that the SNR is 0dB, the 2D-SL0 method basically fails, and the 2D-IFSBL and IFSBL methods provided by the invention can still accurately reconstruct the sparse signals, which shows that the robustness of the sparse signals to noise is better than that of the 2D-SL0 method.
Algorithm performance was further compared by monte carlo experiments. Firstly, the relative errors of the reconstructed sparse signals by the three methods under different observation signal sizes are compared, and experimental parameters are set as follows: m80, K10, SNR 20dB, P, Q, 10 to 60, and 5 steps. Under the condition of each different observed signal size, 100 Monte Carlo experiments are repeatedly carried out, A, B, X and N are randomly valued in each experiment, wherein K-10 non-zero values in the sparse signal X obey standard normal distribution, and the positions in X are completely random. The average sparse reconstruction relative error obtained by the three methods in 100 monte carlo experiments is shown in fig. 4. As can be seen, the performance of the 2D-IFSBL method and the IFSBL method provided by the invention is completely consistent, when P is less than 40, the 2D-SL0 method is better, and when P is more than 40, the relative error obtained by the 2D-IFSBL method and the IFSBL method is lower than that obtained by the 2D-SL0 method.
Further comparing the performances of the three methods under different SNR conditions by Monte Carlo experiments. The experimental parameters were set as follows: M-N-80, P-Q-55, K-10, SNR ranging from-12 dB to 16dB, with a step size of 2 dB. Under each SNR condition, 100 monte carlo experiments were performed, and the average sparse reconstruction relative error of the three methods was recorded, and the results are shown in fig. 5. As can be seen from the figure, the 2D-IFSBL method and the IFSBL method provided by the invention are comprehensively superior to the 2D-SL0 method, and the low relative error of sparse reconstruction is obtained under any SNR condition, so that the strong robustness is verified.
And finally, comparing the operation efficiency of the three methods. Also using the monte carlo experiment, the experimental parameters were set as follows: SNR 20dB, M N. Changing the value of N, wherein the value range is 50-250, the step length is 10, carrying out 100 Monte Carlo experiments on each N condition, and recording the average calculation time of the three algorithms (the calculation platform: Intel (R) Corei7-8550U @1.8GHz), wherein the result is shown in FIG. 6. As can be seen from the figure, the 2D-SL0 method has the highest operation efficiency, and the second method of the 2D-IFSBL method provided by the invention is the lowest IFSBL method. The IFSBL method has high requirement on the calculation memory because of the large matrix calculation, and when N is more than 150, the calculation memory is overrun and the algorithm is invalid. In contrast, on the premise of keeping the performance consistent with the IFSBL, the 2D-IFSBL method provided by the invention improves the highest operation efficiency by more than 100 times, reduces the requirement on the calculation memory, and still only needs about 5 seconds to reconstruct a sparse signal when N is 250, thereby meeting the actual engineering requirements.
The experimental results show that the invention corrects the traditional IFSBL, so that the traditional IFSBL can directly process two-dimensional data. In the problem of two-dimensional sparse reconstruction, the method provided by the invention has the advantages that on the premise of keeping the performance of the traditional IFSBL method, the operation efficiency is improved by more than 100 times, and the requirement on the calculation memory is obviously reduced. Compared with the 2D-SL0 method, the method disclosed by the invention belongs to a statistical sparse reconstruction method and has the advantage of stronger robustness. Although the algorithm performance is lower than that of the 2D-SL0 method under the condition that the observed data size is small, the algorithm performance is obviously better than that of the 2D-SL0 method under the condition of low SNR, and the method is suitable for application scenes with low SNR.

Claims (4)

1. A two-dimensional inversion-free sparse Bayesian learning fast sparse reconstruction method is characterized by comprising the following steps:
s1 sparse representation modeling is carried out on the two-dimensional sparse reconstruction problem:
reconstructing a two-dimensional sparse signal from incomplete observation data, firstly, establishing a relational expression of the two signals, as shown in the following formula:
Y＝AXBT+N (1)
whereinRespectively representing an observed signal, a left dictionary matrix, a sparse signal, a right dictionary matrix and noise,representing a real number set, P, Q, M, N respectively representing the dimension of the matrix, in the two-dimensional sparse reconstruction problem, the dimension of an observation signal is smaller than that of a sparse signal, namely P is less than M and Q is less than N; because the observation signal and the sparse signal are two-dimensional signals, the sparse signal X cannot be reconstructed from the observation signal Y directly by adopting a traditional sparse reconstruction method, and firstly, the observation signal and the sparse signal need to be vectorized:
y＝Φx+n (2)
wherein y ═ vec (y), x ═ vec (x), n ═ vec (n) respectively represent vectorization of observed signals, sparse signals, and noise, vec (·) represents vectorization of the matrix by column stacking;the representation matrix vectorizes the corresponding dictionary matrix:whereinRepresenting the kronecker product of the matrix;
s2 statistically models the vectorized sparse signal x and the vectorized noise n:
in the statistical modeling process, a Gaussian distribution of a two-layer structure is adopted to carry out statistical modeling on the quantitative sparse signal x: in the first layer, it is assumed that the elements of the vectorized sparse signal x obey a gaussian distribution with a mean value of zero and independent of each other:
where γ represents the inverse variance of the vectorization, xn、γnRespectively indicate toQuantizing the sparse signal x and the nth element of the quantized inverse variance γ;
in the second layer, the vectorized inverse variance γ is assumed to obey the gamma distribution:
wherein a and b respectively represent the shape parameter and the size parameter of the gamma distribution;
the statistical modeling of the quantization noise n is also performed by using a gaussian distribution of a two-layer structure: in the first layer, assuming that the vectorized noise n obeys a zero-mean gaussian distribution, the likelihood function of the vectorized observed signal y also obeys a gaussian distribution:
whereinRepresenting a gaussian distribution, I representing an identity matrix, α representing the inverse of the noise variance;
in the second layer, it is assumed that the inverse α of the noise variance obeys the gamma distribution:
wherein c and d respectively represent the shape parameter and the scale parameter of the gamma distribution;
s3 solves the posterior probabilities of the vectorized sparse signal x, the vectorized inverse variance γ, and the noise inverse variance α:
the IFSBL method is based on a likelihood function of the vectorized observation signal y shown in equation (5) and a prior probability of the vectorized sparse signal x shown in equation (3), and the posterior probability of the vectorized sparse signal x is obtained as follows:
where the desired μ and covariance matrix Σ is given by:
whereinAs an auxiliary variable, the number of variables,<·>expressing an expected operator, wherein T is a constant and the value of T is slightly larger than a Lipschitz constant; diag (·) denotes the generation of a diagonal matrix using vectors; as can be seen from equation (9), the covariance matrix Σ is a diagonal matrix, and thus the inversion thereof can be obtained by inverting each element, and the operation efficiency is significantly improved; however, the calculation of the large matrix Φ is involved in the formula (8), the operation efficiency is low, and the requirement on the memory is high; aiming at the problem, the invention corrects the expected solving process shown in the formula (8) so that the two-dimensional data can be directly processed to avoid large matrix calculation caused by matrix vectorization; specifically, formula (9) can be substituted for formula (8):
whereinRepresenting the division of the elements of the matrix, 1MN×1A vector having a size of MN × 1 and all elements of 1 is represented; further will beSubstituting formula (10) and utilizing the relevant properties of the kronecker product, one can obtain:
the above equation can be transformed into a matrix form:
whereinIn the form of matrices of the desired mu, the auxiliary variable z and the inverse of the vectorized variance gamma, 1M×NIs represented by 1MN×1In the form of a matrix;
the posterior probability of the matrix form Γ of the vectorized inverse variance γ is further calculated, which is obtained by means of the bayesian formula, which also obeys the gamma distribution:
wherein the shape parameterAnd scale parameterAs shown in the following formula:
wherein Xm,nThe m-th row and n-th column elements representing the sparse signal X,is expected toThe calculation expression of (a) is:wherein U ism,n、Γm,nThe m-th row and n-th column elements respectively represent U and gamma;
and finally, calculating the posterior probability of the inverse noise variance alpha, which is known by a Bayesian formula and also obeys gamma distribution:
wherein the shape parameter of the a posteriori probabilityAnd a scale parameterRespectively as follows:
wherein sum (·), "indicates the addition of each element in the matrix and the multiplication of each element in the two matrices, respectively;
the desired < α > and < Γ > involved in equations (12), (18) may be obtained from the posterior probabilities shown in equations (16) and (13), respectively, as shown in the following equations:
<Γm,n>matrix form representing vectorized inverse variance γ expectation<Γ>The expectations of the respective elements in (1);
s4 updates the matrix form Z of the auxiliary variable:
after the posterior probabilities of the vectorized sparse signal x, the vectorized inverse variance γ and the noise inverse variance α are calculated, the matrix form Z of the auxiliary variable is further updated by the following formula:
Z＝U (21)
in summary, the process flow of the present invention can be summarized as follows: and (3) circularly iterating the formula (12), the formula (19), the formula (20) and the formula (21) until the relative error between the U obtained by two adjacent iterations reaches a rated threshold, and obtaining the U in the last step, namely the reconstructed sparse signal.
2. The two-dimensional inversion-free sparse Bayesian learning rapid sparse reconstruction method according to claim 1, characterized in that: the parameters are initialized before the iterative process, in particular: u is initialized toWhereinRepresenting a pseudo-inverse of the matrix; the inverse of the noise variance, α, is initialized to α ═ 0.1var (y)-1Wherein var (·) represents a variance operator; the parameters a, b, c, and d in the gamma distributions shown in equations (4) and (6) are initialized to a, b, c, d, and 10-6(ii) a The vectorization inverse variance matrix form gamma is initialized to be 1M×N。
3. The two-dimensional inversion-free sparse Bayesian learning fast sparse reconstruction method according to claim 1 or 2, characterized by: the specific value of the constant T in S3 is: t2 lambdamax(ΦTΦ)+10-10。
4. The two-dimensional inversion-free sparse Bayesian learning fast sparse weight method according to claim 1 orThe method is characterized in that: the rated threshold in S4 is | | | U(k+1)-U(k)||2||U(k)||2＜10-3Wherein U is(k)、U(k+1)Respectively representing the U obtained by the k < th > and k +1 < th > iterations, | | · | | luminous flux2Representing the 2 norm of the matrix.
CN201910854202.5A 2019-09-10 2019-09-10 Two-dimensional inversion-free sparse Bayesian learning rapid sparse reconstruction method Active CN110596645B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201910854202.5A CN110596645B (en) 2019-09-10 2019-09-10 Two-dimensional inversion-free sparse Bayesian learning rapid sparse reconstruction method

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201910854202.5A CN110596645B (en) 2019-09-10 2019-09-10 Two-dimensional inversion-free sparse Bayesian learning rapid sparse reconstruction method

Publications (2)

Publication Number Publication Date
CN110596645A true CN110596645A (en) 2019-12-20
CN110596645B CN110596645B (en) 2021-04-23

Family

ID=68858626

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201910854202.5A Active CN110596645B (en) 2019-09-10 2019-09-10 Two-dimensional inversion-free sparse Bayesian learning rapid sparse reconstruction method

Country Status (1)

Country Link
CN (1) CN110596645B (en)

Cited By (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113030964A (en) * 2021-02-04 2021-06-25 中国人民解放军陆军工程大学 Bistatic ISAR (inverse synthetic aperture radar) thin-aperture high-resolution imaging method based on complex Laplace prior
CN113466864A (en) * 2021-08-16 2021-10-01 中国人民解放军空军工程大学 Fast joint inverse-free sparse Bayesian learning super-resolution ISAR imaging algorithm
CN113900095A (en) * 2021-08-24 2022-01-07 西安电子科技大学 Metamaterial dynamic grid-based inverse-free sparse Bayesian imaging method and system
CN114280533A (en) * 2021-12-23 2022-04-05 哈尔滨工程大学 Sparse Bayesian DOA estimation method based on l0 norm constraint

Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN103888145A (en) * 2014-03-28 2014-06-25 电子科技大学 Method for reconstructing signals
CN104767535A (en) * 2015-03-31 2015-07-08 电子科技大学 Low-complexity massive sparse signal reconstruction method
US9791563B1 (en) * 2014-01-08 2017-10-17 National Technology & Engineering Solutions Of Sandia, Llc Joint synthetic aperture radar plus ground moving target indicator from single-channel radar using compressive sensing
CN108226927A (en) * 2017-12-14 2018-06-29 电子科技大学 SAR imaging methods based on weighted iteration minimum sparse Bayesian restructing algorithm
CN109696671A (en) * 2018-12-29 2019-04-30 南京汇君半导体科技有限公司 High-resolution through-wall radar imaging method based on group structure and correlation study
CN110068805A (en) * 2019-05-05 2019-07-30 中国人民解放军国防科技大学 High-speed target HRRP reconstruction method based on variational Bayesian inference

Patent Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US9791563B1 (en) * 2014-01-08 2017-10-17 National Technology & Engineering Solutions Of Sandia, Llc Joint synthetic aperture radar plus ground moving target indicator from single-channel radar using compressive sensing
CN103888145A (en) * 2014-03-28 2014-06-25 电子科技大学 Method for reconstructing signals
CN104767535A (en) * 2015-03-31 2015-07-08 电子科技大学 Low-complexity massive sparse signal reconstruction method
CN108226927A (en) * 2017-12-14 2018-06-29 电子科技大学 SAR imaging methods based on weighted iteration minimum sparse Bayesian restructing algorithm
CN109696671A (en) * 2018-12-29 2019-04-30 南京汇君半导体科技有限公司 High-resolution through-wall radar imaging method based on group structure and correlation study
CN110068805A (en) * 2019-05-05 2019-07-30 中国人民解放军国防科技大学 High-speed target HRRP reconstruction method based on variational Bayesian inference

Non-Patent Citations (3)

* Cited by examiner, † Cited by third party
Title
HUIPING DUAN ET AL.: "Fast Inverse-Free Sparse Bayesian Learning via Relaxed Evidence Lower Bound Maximization", 《IEEE SIGNAL PROCESSING LETTERS》 *
XINGYU HE ET AL.: "Superresolution radar imaging based", 《JOURNAL OF APPLIED REMOTE SENSING》 *
胡仁荣 等: "模式耦合快速求逆稀疏贝叶斯MIMO雷达成像", 《弹箭与制导学报》 *

Cited By (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113030964A (en) * 2021-02-04 2021-06-25 中国人民解放军陆军工程大学 Bistatic ISAR (inverse synthetic aperture radar) thin-aperture high-resolution imaging method based on complex Laplace prior
CN113466864A (en) * 2021-08-16 2021-10-01 中国人民解放军空军工程大学 Fast joint inverse-free sparse Bayesian learning super-resolution ISAR imaging algorithm
CN113466864B (en) * 2021-08-16 2023-07-04 中国人民解放军空军工程大学 Rapid combined inverse-free sparse Bayes learning super-resolution ISAR imaging algorithm
CN113900095A (en) * 2021-08-24 2022-01-07 西安电子科技大学 Metamaterial dynamic grid-based inverse-free sparse Bayesian imaging method and system
CN113900095B (en) * 2021-08-24 2024-04-26 西安电子科技大学 Inverse sparse Bayesian imaging method and system based on metamaterial dynamic grid
CN114280533A (en) * 2021-12-23 2022-04-05 哈尔滨工程大学 Sparse Bayesian DOA estimation method based on l0 norm constraint

Also Published As

Publication number Publication date
CN110596645B (en) 2021-04-23

Similar Documents

Publication Publication Date Title
CN110596645B (en) Two-dimensional inversion-free sparse Bayesian learning rapid sparse reconstruction method
US10985777B2 (en) Signal recovery via deep convolutional networks
US20220300819A1 (en) System and method for designing efficient super resolution deep convolutional neural networks by cascade network training, cascade network trimming, and dilated convolutions
Liu et al. Group fisher pruning for practical network compression
Zheng et al. Implicit dual-domain convolutional network for robust color image compression artifact reduction
WO2018039904A1 (en) Block sparse compressive sensing based infrared image reconstruction method and system thereof
US11250600B2 (en) Method for processing X-ray computed tomography image using neural network and apparatus therefor
CN113177882B (en) Single-frame image super-resolution processing method based on diffusion model
Ziabari et al. 2.5 D deep learning for CT image reconstruction using a multi-GPU implementation
WO2020048354A1 (en) Neural network model compression method and apparatus, and computer device
CN113298734B (en) Image restoration method and system based on mixed hole convolution
CN106097278A (en) The sparse model of a kind of multidimensional signal, method for reconstructing and dictionary training method
US20190355093A1 (en) Image processing apparatus using neural network and method performed by image processing apparatus
US20240029204A1 (en) Image processing method, system, device and storage medium
WO2022151586A1 (en) Adversarial registration method and apparatus, computer device and storage medium
Cao et al. CS-MRI reconstruction based on analysis dictionary learning and manifold structure regularization
CN114118406A (en) Quantitative compression method of convolutional neural network
CN111369433A (en) Three-dimensional image super-resolution reconstruction method based on separable convolution and attention
Liu et al. Computation-performance optimization of convolutional neural networks with redundant filter removal
CN112288046B (en) Mixed granularity-based joint sparse method for neural network
Hu et al. Patch-based low-rank minimization for image denoising
CN111798531B (en) Image depth convolution compressed sensing reconstruction method applied to plant monitoring
CN113673419A (en) Beam domain fast sparse Bayesian direction estimation method suitable for strong interference environment
Danielyan et al. Deblurring of Poissonian images using BM3D frames
CN110174657B (en) Direction-of-arrival estimation method based on rank-one dimension reduction model and block matrix recovery

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
GR01 Patent grant
GR01 Patent grant